OpenAI Responses

Connect AgentCompass to OpenAI Responses-compatible endpoints.

Use the openai-responses protocol for endpoints compatible with the OpenAI Responses API, commonly exposed at /v1/responses.

Configure the Protocol

export MODEL_NAME=""
export MODEL_BASE_URL="https://your-endpoint.example/v1"
export MODEL_API_KEY="sk-..."

agentcompass run <benchmark> <harness> "$MODEL_NAME" \
  --env <environment> \
  --model-base-url "$MODEL_BASE_URL" \
  --model-api-key "$MODEL_API_KEY" \
  --model-api-protocol openai-responses \
  --model-params '{"reasoning":{"effort":"high"}}'

Compatibility Conversion

AgentCompass converts chat-style message history and function tools into Responses input items when a runtime component uses the shared protocol client. It also normalizes Responses output into the chat-like internal shape expected by common harness logic.

For compatibility with existing model configurations:

• max_tokens becomes max_output_tokens unless max_output_tokens is already set;
• reasoning_effort becomes reasoning.effort unless reasoning is already set;
• explicit Responses-native fields pass through unchanged when supported by the selected harness and endpoint.

When to Use It

Choose openai-responses when the provider and selected harness explicitly support the Responses API, especially for reasoning or stateful tool-call workflows. Do not select it only because an endpoint is OpenAI-compatible; many compatible providers implement Chat Completions but not Responses.

See OpenAI Chat for Chat Completions endpoints and Choose a Harness for protocol compatibility.